Symptom

$ tools/run_pascal_conformance.sh ./compiler/pascal26 library_candidates/fpc-testsuite/tests/test --only 'tgeneric*'
test-pascal-conformance: 10 pass, 51 fail, 43 skip, 3 auto-gated (of 107)

$ tools/run_pascal_conformance.sh /home/neo/frank1/compiler/pascal26 \
      /home/neo/frank1/library_candidates/fpc-testsuite/tests/test --only 'tgeneric*'
test-pascal-conformance: 61 pass, 0 fail, 43 skip, 3 auto-gated (of 107)

The per-test output names the real cause, but only if you look at one:

FAIL tgeneric12.pp — compile error:
    timeout: failed to execute process: No such file or directory (os error 2)

Cause

Both compile paths run the compiler from inside another directory:

( cd "$SUITE"  && timeout "$TIMEOUT_S" "$CC" $CCFLAGS "$name" "$bin" ) ...
( cd "$WORK"   && timeout "$TIMEOUT_S" "$CC" $CCFLAGS "drv_$uname.pas" "$bin" ) ...

$CC defaults to $ROOT/compiler/pascal26 (absolute), so the default invocation is fine and testmgr's is fine. Only an operator passing the path by hand hits it — which is exactly the debugging position where a wall of red is most expensive, because it reads as "my change broke 51 tests".

$SUITE has the same shape and the same fix.

Fix

Absolutise both at parse time, right after the positional arguments are taken:

case "$CC"    in /*) ;; *) CC="$(CDPATH= cd -- "$(dirname -- "$CC")" && pwd)/$(basename -- "$CC")" ;; esac
case "$SUITE" in /*) ;; *) SUITE="$(CDPATH= cd -- "$SUITE" && pwd)" ;; esac

and, separately, make a compiler that cannot be EXECUTED a hard error rather than a per-test compile failure: the [ -x "$CC" ] check already exists but runs before the cd, so it passes on a relative path that will not resolve later.

tools/run_c_conformance.sh should be checked for the same shape (it is the file this one was written from).

Fixed 2026-08-26 (Track T)

Both paths are absolutised immediately after the positional arguments are taken, before anything can cd, and the [ -x "$CC" ] check moved to after that — which is what makes it meaningful. It used to run against the path as given, so a relative path that would not resolve post-cd passed the check and then failed 51 times, one test at a time. A compiler that cannot be executed is now a hard error: the difference between "your setup is wrong" and "your compiler is broken" is exactly what this ticket is about.

Verified, same command as the symptom above:

Relative and absolute now agree exactly, which is the property that was missing.

tools/run_c_conformance.sh does NOT have this bug — checked, because this file was written from it and the ticket asked. It takes $CC and $SUITE the same way, but it never cds, so a relative path keeps resolving and the failure cannot occur. Same argument shape, no manifestation. Left alone rather than "fixed" symmetrically: an edit that changes no behaviour still has to be read and understood by the next person, and a comment claiming to fix a bug that was never there is worse than no comment.

If a cd is ever added there, this is the trap to remember — which is why that sentence is in this ticket rather than in a comment nobody will find.

Log

• 2026-08-26 — resolved, commit bc0ffebf9.